3D Reconstruction of cellular images from microfabricated imagers using fully-adaptive deep neural networks

Millimeter-scale multi-cellular level imagers enable various applications, ranging from intraoperative surgical navigation to implantable sensors. However, the tradeoffs for miniaturization compromise resolution, making extracting 3D cell locations challenging-critical for tumor margin assessment and therapy monitoring. This work presents three machine-learning-based modules that extract spatial information from single image acquisitions using custom-made millimeter-scale imagers. The neural networks were trained on synthetically-generated (using Perlin noise) cell images. The first network is a convolutional neural network estimating the depth of a single layer of cells, the second is a deblurring module correcting for the point spread function (PSF). The final module extracts spatial information from a single image acquisition of a 3D specimen and reconstructs cross-sections, by providing a layered "map" of cell locations. The maximum depth error of the first module is 100 µm, with 87% test accuracy. The second module's PSF correction achieves a least-square-error of only 4%. The third module generates a binary "cell" or "no cell" per-pixel labeling with an accuracy ranging from 89% to 85%. This work demonstrates the synergy between ultra-small silicon-based imagers that enable in vivo imaging but face a trade-off in spatial resolution, and the processing power of neural networks to achieve enhancements beyond conventional linear optimization techniques.